Cheapest way to run gpt-oss-120b locally
gpt-oss-120b's Q4_K_M GGUF is ~62.8 GB (bartowski, Hugging Face), so it needs ~64 GB+ of VRAM to run fully on GPU, or partial CPU offload (much slower, not estimated here).
gpt-oss-120b needs about 64.3 GB of VRAM. The Q4_K_M GGUF is 62.84 GB (source), plus our 1.5 GB allowance for the runtime and a short-context KV cache.
No GPU or system with a dated price has that much memory. The 2 devices that fit it (Apple Mac Studio M5 Ultra (64-core GPU, 96 GB), Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB) have no verified price in our data.
None of our 5 fully priced builds reach 64.3 GB — the largest is 48 GB: dual used RTX 3090 at 48 GB. Cheapest multi-GPU route: 3× AMD Radeon RX 7900 XTX = $4,050 for 72 GB, ~266 tok/s (cards only, no other parts).
Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "check price" means we have no verified price.
Every GPU and system that fits gpt-oss-120b
| GPU or system | VRAM | Bandwidth | Speed estimate | Low price | $ per tok/s |
|---|---|---|---|---|---|
| Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) | 96 GB | 1200 GB/s | ~332 tok/s | check price | — |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | 273 GB/s | ~76 tok/s | check price | — |
Reaching 64.3 GB with identical cards
| Cards | Total VRAM | Cards only | Speed estimate | $ per tok/s (cards only) |
|---|---|---|---|---|
| 3× AMD Radeon RX 7900 XTX | 72 GB | $4,050 | ~266 tok/s | $15 |
| 3× NVIDIA GeForce RTX 3090 | 72 GB | $5,667 | ~259 tok/s | $22 |
| 3× NVIDIA GeForce RTX 4090 | 72 GB | $9,600 | ~280 tok/s | $34 |
| 3× NVIDIA GeForce RTX 5090 | 96 GB | $19,500 | ~496 tok/s | $39 |
Card prices only — no CPU, board, memory, PSU or case. Speed uses one card's bandwidth because a layer split reads one card at a time.
Our fully priced builds
| Build | VRAM | Total | Fits gpt-oss-120b? | Speed estimate | $ per tok/s |
|---|---|---|---|---|---|
| 16 GB starter: Radeon RX 9070 | 16 GB | $1,744.44 | ❌ | — | — |
| 16 GB NVIDIA: RTX 5060 Ti 16 GB | 16 GB | $1,834.44 | ❌ | — | — |
| 24 GB: one used RTX 3090 | 24 GB | $2,933.45 | ❌ | — | — |
| 32 GB: two Radeon RX 9070 | 32 GB | $3,065.92 | ❌ | — | — |
| 48 GB: dual used RTX 3090 | 48 GB | $5,443.94 | ❌ | — | — |
Speed calculator estimate
Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works
Questions people ask
How much VRAM do you need for gpt-oss-120b?
About 64.3 GB. The Q4_K_M GGUF is 62.84 GB and we add 1.5 GB for the runtime and a short-context KV cache. 2 of the 36 GPUs and systems we track have that much memory. Long contexts need more than our allowance.
What is the cheapest way to run gpt-oss-120b?
No GPU or system with a dated price has that much memory: the 2 devices that fit it (Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) and Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB) have no verified price in our data. None of our 5 fully priced builds reach 64.3 GB — the largest, the 48 GB: dual used RTX 3090 build, tops out at 48 GB. The cheapest way to reach the memory is to split the model across identical cards: 3 x AMD Radeon RX 7900 XTX is $4,050 for 72 GB, at an estimated 266 tok/s — cards only, before the CPU, board, memory, PSU and case.
Can you run gpt-oss-120b on multiple GPUs?
Yes: VRAM adds up across cards, so 3 x AMD Radeon RX 7900 XTX (72 GB, $4,050) covers the 64.3 GB this model needs. A layer split still reads from one card at a time, so the estimate stays at one card's bandwidth: about 266 tok/s.
How fast does gpt-oss-120b run?
Estimates across the 2 devices that fit it range from about 76 tok/s on the Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB (273 GB/s) to about 332 tok/s on the Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) (1200 GB/s). These are bandwidth estimates, not benchmarks, and they account for only 5.1B of 116.83B parameters being active per token.
Related
- What the AMD Radeon RX 7900 XTX can run
- What the NVIDIA GeForce RTX 3090 can run
- What the Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) can run
- What the Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB can run
- What GPU you need for gpt-oss-120b
- What GPU you need for Qwen3-32B
- Can the Radeon RX 7900 XTX run Qwen2.5-Coder-32B-Instruct?
- Can the Radeon RX 7900 XTX run Qwen3-32B?
- Can the GeForce RTX 3090 run Qwen2.5-Coder-32B-Instruct?
- Can the GeForce RTX 3090 run Qwen3-32B?
- Build: 48 GB: dual used RTX 3090 — $5,443.94, 48 GB
- Guide: RTX 3090 vs RTX 5090 for local LLMs
- Guide: Strix Halo vs a GPU for local AI
- Guide: Mac vs PC for local LLMs